Phase 0: Configuration Wizard
  ├─ Detect escape hatches ("using design doc <path>", "using impl plan <path>")
  ├─ Clarify feature essence (1-2 sentences)
  └─ Collect preferences: autonomous_mode, parallelization, worktree, post_impl
    ↓
Phase 1: Research
  ├─ Generate codebase questions from feature request
  ├─ Dispatch research subagent (codebase, web, MCP, user links)
  ├─ Extract ambiguities (MEDIUM/LOW/UNKNOWN confidence)
  └─ GATE: Research Quality Score = 100% (coverage, evidence, ambiguity resolution)
    ↓
Phase 1.5: Informed Discovery
  ├─ Resolve ambiguities via AskUserQuestion (ARH pattern for response handling)
  ├─ Generate 7-category discovery questions (informed by research)
  ├─ Build glossary, synthesize design_context
  ├─ Create Understanding Document
  ├─ GATE: Completeness Score = 100% (11 validation functions)
  └─ Invoke devils-advocate skill for adversarial review
    ↓
Phase 2: Design (skip if escape hatch)
  ├─ Subagent invokes brainstorming (SYNTHESIS MODE - no questions)
  ├─ Subagent invokes design-doc-reviewer
  ├─ GATE: User approval (interactive) or auto-proceed (autonomous)
  └─ Subagent invokes executing-plans to fix findings
    ↓
Phase 3: Implementation Planning (skip if impl plan escape hatch)
  ├─ Subagent invokes writing-plans
  ├─ Subagent invokes implementation-plan-reviewer
  ├─ GATE: User approval per mode
  ├─ Subagent invokes executing-plans to fix
  ├─ Execution mode analysis (tokens/tasks/tracks → swarmed|delegated|direct)
  ├─ If swarmed: Generate work packets, spawn sessions, EXIT
  └─ If delegated/direct: Continue to Phase 4
    ↓
Phase 4: Implementation
  ├─ Setup worktree(s) per preference
  ├─ If parallel worktrees: Complete setup/skeleton FIRST, commit, then create
  ├─ For each task:
  │   ├─ Subagent invokes test-driven-development
  │   ├─ Completion verification (trace acceptance criteria through code)
  │   ├─ Subagent invokes requesting-code-review
  │   └─ Subagent invokes fact-checking
  ├─ If parallel worktrees: Subagent invokes worktree-merge
  ├─ Comprehensive implementation audit (all tasks, integrations, traceability)
  ├─ Run test suite (invoke systematic-debugging if failures)
  ├─ Subagent invokes audit-green-mirage
  ├─ Comprehensive fact-checking
  ├─ Pre-PR fact-checking
  └─ Subagent invokes finishing-a-development-branch per post_impl preference

SESSION_PREFERENCES = {
  autonomous_mode: "autonomous"|"interactive"|"mostly_autonomous",
  parallelization: "maximize"|"conservative"|"ask",
  worktree: "single"|"per_parallel_track"|"none",
  post_impl: "offer_options"|"auto_pr"|"stop",
  escape_hatch: null | {type, path, handling}
}

SESSION_CONTEXT = {
  feature_essence: {},
  research_findings: {},
  design_context: {}  // Comprehensive context passed to all subagents
}

IF request contains ["refactor", "reorganize", "extract", "migrate", "split", "consolidate"]:
  refactoring_mode = true
  SESSION_PREFERENCES.refactoring_mode = true

1. Identify boundary (what calls in, what calls out)
2. Create abstraction at boundary (interface/facade)
3. Route existing code through abstraction
4. Verify behavior unchanged (tests pass)
5. Implement new version behind abstraction
6. Gradually shift traffic (feature flag or config)
7. Monitor for behavior differences
8. Remove old implementation when confident

[ ] Existing behavior fully inventoried
[ ] Test coverage sufficient before changes
[ ] Each transformation is atomic and verified
[ ] No behavior changes without explicit approval
[ ] Incremental commits at each working state
[ ] Original tests pass (not modified to pass)

Task: "[Phase] [action]"
prompt: |
  First, invoke the [skill-name] skill using the Skill tool.
  Then follow its complete workflow.

  ## Context for the Skill
  [Only context the skill needs - no duplicated instructions]

IF tasks > 25 OR context_usage > 80%: swarmed (spawn sessions, EXIT)
IF context_usage > 65% OR (tasks > 15 AND tracks >= 3): swarmed
IF tasks > 10 OR context_usage > 40%: delegated (heavy subagent use)
ELSE: direct (minimal delegation)